Loading ToolBench
Extended MemPalace fork with document ingestion, image assets, MCP multimodal output, and folder watch. Provides semantic memory storage and retrieval via ChromaDB with palace metaphor (wings/rooms/drawers).
MemPalace has 15 tools with basic schemas and descriptions, but significant quality gaps prevent production readiness. No error handling guidance, no recovery paths, and no input validation hints. The server exposes internal implementation details (drawer_id, wing/room names) without explaining how to obtain them. Critical tools like mempalace_add_drawer and mempalace_delete_drawer lack confirmation/dry-run patterns for destructive operations.
file verbatim content into a wing/room
check if content already exists before filing
remove a drawer by ID
Record what happened, what you learned, what matters in this session
Return the AAAK dialect specification
Fetch stored image assets by filename or figure reference
full wing → room → count tree
Add a fact to the knowledge graph
Destructive tools (mempalace_delete_drawer, mempalace_kg_invalidate) lack confirmation/dry-run patterns. Agents can delete data without verification.
Output schemas are undocumented. Tool descriptions do not specify what fields are returned, forcing LLMs to guess structure and plan downstream calls blindly.
Parameter descriptions are missing or vague. 'wing' and 'room' parameters lack guidance on valid values, format, or how to discover them. LLMs cannot infer valid inputs.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 41 | <=2025-11-25 | v2 |
Mark a fact as invalid in the knowledge graph
Query the knowledge graph for facts about entities
rooms within a wing
all wings with drawer counts
semantic search, optional wing/room filter
total drawers, wing/room breakdown
Walk the palace graph from a room. Find connected ideas across wings.
No error handling guidance. Tools return errors without recovery hints. E.g., 'drawer_id not found' should suggest 'Call mempalace_search() to find the drawer ID first.'
Tool descriptions are too short (10-25 chars for mempalace_list_wings, mempalace_get_taxonomy, mempalace_delete_drawer). Rubric baseline is 34-392 chars. LLMs cannot determine when to use these tools.